graphics: route compact color exports through CB_SHADER_MASK

This commit is contained in:
Claxten
2026-10-07 18:22:46 +04:00
committed by nmzik
parent 3ba6ebae81
commit 21a1346e27
7 changed files with 92 additions and 35 deletions
@@ -589,8 +589,12 @@ static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
const auto& sh_regs = ctx.GetShaderRegisters();
const auto write_mask = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
uint32_t output_mask = 0;
for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
if (sh_regs.target_output_mode[slot] != 0 &&
for (uint32_t index = 0; index < RENDER_COLOR_ATTACHMENTS_MAX; index++) {
const auto slot = ShaderPixelExportTarget(sh_regs.m_cbShaderMask, index);
if (slot >= RENDER_COLOR_ATTACHMENTS_MAX) {
break;
}
if (sh_regs.target_output_mode[index] != 0 &&
render_target_mask_slot(write_mask, slot) != 0) {
output_mask |= 1u << slot;
}
@@ -902,7 +906,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
if (state.ps_active) {
for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt) {
mrt_mask |= 1u << output.index;
mrt_mask |= 1u << output.location;
}
}
}
@@ -390,16 +390,23 @@ void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, Sha
AddOutput(info, StageOutputKind::Parameter, export_info.index,
export_info.index, fmt::format("out_param_{}", export_info.index));
break;
case ExportTargetKind::Mrt:
case ExportTargetKind::Mrt: {
if (alpha_remap && export_info.index != 0) {
break;
}
AddOutput(info, StageOutputKind::Mrt, export_info.index, export_info.index,
const auto slot = input_info.pixel->dual_source_blending
? 0u : ShaderPixelExportTarget(input_info.pixel->target_shader_mask,
export_info.index);
if (slot >= 8) {
break;
}
AddOutput(info, StageOutputKind::Mrt, export_info.index, slot,
fmt::format("out_mrt_{}", export_info.index));
if (alpha_remap) {
AddOutput(info, StageOutputKind::Mrt, 1, 1, "out_mrt_1");
AddOutput(info, StageOutputKind::Mrt, 1, 0, "out_mrt_1");
}
break;
}
default: break;
}
}
+6 -2
View File
@@ -561,10 +561,12 @@ static void ShaderGetStaticInputInfoPS(
ps_info.interpolator_settings[i] = sh.ps_interpolator_settings[i];
}
ps_info.target_shader_mask = sh.m_cbShaderMask;
for (int i = 0; i < 8; i++) {
const auto slot = ShaderPixelExportTarget(ps_info.target_shader_mask, i);
ps_info.target_output_mode[i] = sh.target_output_mode[i];
ps_info.target_export_mapping[i] = sh.target_output_mode[i] != 0
? target_export_mapping[i]
ps_info.target_export_mapping[i] = sh.target_output_mode[i] != 0 && slot < 8
? target_export_mapping[slot]
: Prospero::ColorComponentMapping {};
}
}
@@ -668,6 +670,8 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
key.push_back(static_cast<uint32_t>(info.dual_source_blending));
key.push_back(static_cast<uint32_t>(info.alpha_blend_source));
key.insert(key.end(), std::begin(info.target_output_mode), std::end(info.target_output_mode));
const auto mask = info.target_shader_mask;
key.push_back((mask | (mask >> 1u) | (mask >> 2u) | (mask >> 3u)) & 0x11111111u);
for (uint32_t base = 0; base < info.target_export_mapping.size(); base += 4u) {
uint32_t packed = 0;
for (uint32_t i = 0; i < 4u; i++) {
+2
View File
@@ -170,6 +170,7 @@ struct ShaderPixelInputInfo {
uint32_t ps_perspective_sample_vgpr = UINT32_MAX;
uint32_t ps_perspective_centroid_vgpr = UINT32_MAX;
uint8_t target_output_mode[8] = {};
uint32_t target_shader_mask = UINT32_MAX;
std::array<Prospero::ColorComponentMapping, 8> target_export_mapping = {};
uint32_t scratch_size_dwords = 0;
bool ps_pos_x = false;
@@ -310,6 +311,7 @@ struct ShaderMappedData {
void ShaderInit();
void ShaderMapUserData(uint64_t addr, const ShaderMappedData& data);
uint32_t ShaderPixelExportTarget(uint32_t shader_mask, uint32_t export_index);
void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info);
void ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info);
@@ -11,6 +11,19 @@ constexpr uint32_t PsInputFlatShade = 0x00000400u;
} // namespace
uint32_t ShaderPixelExportTarget(uint32_t shader_mask, uint32_t export_index) {
for (uint32_t target = 0; target < 8u; target++) {
if (((shader_mask >> (target * 4u)) & 0xfu) == 0u) {
continue;
}
if (export_index == 0u) {
return target;
}
export_index--;
}
return UINT32_MAX;
}
uint32_t ShaderPixelParameterMappedLocation(const ShaderPixelInputInfo& info, uint32_t input) {
return input < info.input_num ? info.interpolator_settings[input] & PsInputOffsetMask : input;
}